Browsing by Subject "Deep learning"
Now showing 1 - 2 of 2
Results Per Page
Sort Options
- ItemOpen AccessApplication of probabilistic deep learning models to simulate thermal power plant processes(2022) Raidoo, Renita Anand; Laubscher, RynoDeep learning has gained traction in thermal engineering due to its applications to process simulations, the deeper insights it can provide and its abilities to circumvent the shortcomings of classic thermodynamic simulation approaches by capturing complex inter-dependencies. This works sets out to apply probabilistic deep learning to power plant operations using historic plant data. The first study presented, entails the development of a steady-state mixture density network (MDN) capable of predicting effective heat transfer coefficients (HTC) for the various heat exchanger components inside a utility scale boiler. Selected directly controllable input features, including the excess air ratio, steam temperatures, flow rates and pressures are used to predict the HTCs. In the second case study, an encoder-decoder mixturedensity network (MDN) is developed using recurrent neural networks (RNN) for the prediction of utility-scale air-cooled condenser (ACC) backpressure. The effects of ambient conditions and plant operating parameters, such as extraction flow rate, on ACC performance is investigated. In both case studies, hyperparameter searches are done to determine the best performing architectures for these models. Comparisons are drawn between the MDN model versus standard model architecture in both case studies. The HTC predictor model achieved 90% accuracy which equates to an average error of 4.89 W m2K across all heat exchangers. The resultant time-series ACC model achieved an average error of 3.14 kPa, which translate into a model accuracy of 82%.
- ItemOpen AccessLeveraging artificial intelligence for enhanced echocardiography: using machine learning for Congenital Heart Disease(2026) Dzonzi, Brenald; Ngoepe, Malebogo; Mishra, AmitCoarctation of the aorta (CoA) is a congenital cardiovascular defect characterised by a focal narrowing of the descending aorta, resulting in elevated cardiac afterload and often necessitating lifelong haemodynamic surveillance. Although computational fluid dynamics (CFD) enables non-invasive assessment of pressure gradients and flow fields, its reliance on labour-intensive image segmentation, mesh generation, and substantial compute time limits its applicability for rapid parameter studies and routine clinical use, particularly in resource-constrained settings. Neural network surrogates offer a promising alternative by learning input–output mappings from CFD data at a fraction of the inference cost, yet the relative merits of physics informed versus purely data-driven architectures remain insufficiently characterised for patient-specific cardiovascular geometries. This dissertation benchmarks physics-informed neural networks (PINNs) against data-only deep neural networks (DNNs) as mesh-free surrogates for haemodynamic prediction in CoA, with evaluation on a patient-specific geometry from the PROTEA project at the University of Cape Town. A structured benchmark ladder of increasing physical complexity is employed: (i) analytic ordinary and partial differential equations (spring–mass system, two-dimensional Poisson equation) providing exact reference solutions; (ii) canonical laminar flows (planar Poiseuille and three-dimensional pipe flow); (iii) a parametric stenosis series with varying constriction severity; and (iv) a clinically derived CoA geometry with pulsatile inflow. Model families—supervised DNNs with optional random Fourier feature embeddings, physics-only PINNs, hybrid PINNs combining data and residual losses, and a signed-distance-function-augmented variant—are implemented in PyTorch and trained using automatic mixed precision, adaptive loss weighting, and reproducible multi-seed protocols. OpenFOAM supplies the ground-truth CFD datasets and boundary conditions throughout. Results demonstrate that physics constraints confer a substantial advantage under extreme data scarcity (the PINN achieves an 8.5× lower extrapolation error than the DNN at ten labelled points on the Poisson benchmark), but this benefit diminishes as labelled data increase. On the patient-specific CoA geometry, the baseline DNN and Fourier-enhanced DNN satisfy the adopted pressure-drop acceptance criterion (below 1% mean relative error) at both 50% and 100% data fractions, whereas the hybrid PINN exhibits systematic gradient imbalance between its data and physics objectives. Fourier feature embeddings deliver comparable accuracy to the baseline at one to two orders of magnitude lower training cost, representing the most favourable cost–accuracy trade-off observed. All models are validated against CFD rather than in-vivo measurements; clinical deployment would require subsequent catheterisation or four-dimensional flow MRI validation and multi-patient generalisation. This work establishes reproducible technical foundations—encompassing data generation, training, and post-processing pipelines—for neural-network-based haemodynamic surrogates in congenital heart disease, contributing toward the long-term goal of accessible, rapid cardiovascular assessment.